AI consulting for small businesses: what it should include and what it costs

Finds the P&L problem first. A good consultant does not arrive with a tool. They walk your operation, find the work people do by hand every day, and put a number on it: hours, days sales outstanding, error cost, missed calls. If they cannot name the line on your P&L a project moves, the project is a hobby.

Ranks the use cases. There are always more candidates than budget. Ranking by value and risk, in writing, is the deliverable. It tells you what to build first and why, and it is useful even if you never hire the consultant again.

Builds inside your software. Your accounting system, your dispatch or field-service tool, your inbox, your phone. AI that lives in a separate app nobody opens is not adopted. AI that reads the invoice in your inbox and puts it in your approval queue in QuickBooks is.

Hands you the keys. Source code, data, accounts. If the consultant disappears, your system keeps running. If they will not agree to that in writing, keep looking.

Stays to run it. Vendors change APIs; portals change pages. Somebody has to be accountable, with a response time. Ask who.

1. Which line on my P&L does this project move, and by roughly how much?
2. Who does the work, and can I talk to them before I sign?
3. Where does the software run, and whose name is on the account?
4. Who owns the code and the data when we are done?
5. What happens when a vendor changes an API? Who fixes it, and how fast, in writing?
6. Where does a person stay in the loop, and how do I see what the system did?
7. What data of mine goes to which AI provider, and under what terms?
8. What does "done" mean, in numbers, and what happens if it misses?
9. What is the fixed price, and what would change it?
10. What would you tell me not to build?

A good consultant has answers to all ten and is glad you asked. The last one is the tell.

A short assessment with the people who do the work, not just the owner. A ranked list with numbers. One build, small enough to finish in weeks, with success criteria written before it starts and a shadow period before it replaces the manual process. A handoff of code, data, and accounts. Then a decision, with real numbers in hand, about the next one.

What it should not look like: a six-month roadmap, a platform migration, a "center of excellence," or anything that requires you to hire before you see a result.

Market figures below are from our survey of published rates in September 2026 and will drift. Treat them as ranges.

Assessments. A paid AI assessment or roadmap from a consultancy runs roughly $3,500 to $15,000 depending on depth and whether it includes a build. Free assessments exist; they are sales calls with a checklist. Ours starts at $10,000, includes two days on site, a scored backlog you keep, and the first use case built, and it is credited in full against the first workstream.

Hourly. The going AI consultant hourly rate is roughly $150 to $500, with freelancers at the low end and firms at the high end. Hourly works for advice. It works badly for building, because the incentive runs the wrong way: the longer it takes, the more they earn. We quote builds as a fixed fee, $10,000 to $70,000 per workstream, before work starts, so the number you approve is the number you pay.

Managed plans. After go-live, expect $600 to $2,200 a month for monitoring, fixes when a vendor changes something, and a bank of engineering hours. Cheaper than that usually means nobody is watching. Prices verified September 2026. Our full pricing →

For an assessment: a written, ranked list of use cases with a P&L line and a risk note on each, a systems inventory, a recommended first build with success criteria, and governance notes on where a person stays in the loop and what data needs care. If it is a build, the software running in production with a shadow period behind it, the code in your repository, accounts in your name, and a runbook. For a managed plan: monitoring you can see, a monthly report of what ran and what broke, a bank of hours, and a response time in writing.

If a proposal does not name those deliverables, ask what you are buying.

Most of the cost of a good AI build is not the model. Model usage for a back-office agent at a company your size is usually tens of dollars a month. The cost is the integration work, connecting to your accounting system and your dispatch tool and handling their quirks, and the judgment work of deciding what the agent may do on its own and what it must hand to a person. A consultant who quotes mostly for "AI" and little for integration has not looked at your systems yet.

Guaranteed ROI. Nobody can guarantee a return before they have seen your data. A consultant who promises one is selling.

Platform lock-in. If the solution only works on the consultant's platform, you are renting, and the rent goes up. Ask who owns the code and where it runs.

Junior staff after the sale. The person who sold you may not be the person who builds. Ask who does the work and whether you can talk to them.

No owner of the outcome. If the engagement ends at "recommendations," nobody is accountable for whether anything changed. Insist on a build, with success criteria written down, or at least a named person who owns the result.

AI for its own sake. If the first conversation is about models and not about your dispatch board, walk.

Do not start with a strategy. Start with one job a person does by hand every day, with a number attached, and get it running. Then the next one.

Our version of this is a two-day AI Readiness Assessment: day one with the people who do the work, day two scoring every candidate on P&L impact and risk, then the top item built and in production in two weeks. You keep the ranked backlog either way, and you don't pay for the build if it misses the agreed mark. How the assessment works →

An AI agent is software that does a job, not software that answers a question. Three that pay back fastest for companies without a big back office:

AP invoice capture. Reads invoices from email, matches to the PO or vendor history, codes them, and queues them for approval in QuickBooks. P&L line: bookkeeper hours, duplicate payments, early-pay discounts.

Phone receptionist wired into scheduling. Answers every call, qualifies, books into your scheduling or dispatch tool, transfers emergencies to a person. P&L line: jobs booked from calls you used to miss.

The weekly owner report. Revenue, margin, cash, AR aging, from every system into one page on Monday. P&L line: decisions made a week earlier.

The twelve agents we build most, with systems and P&L lines →